Knowledge cutoff
The rough date where a model's built-in training knowledge stops being current. Newer facts need live sources, not confident guessing.
See it
What it is
A knowledge cutoff is the rough latest date represented in a model's training knowledge. It is not a clean snapshot: the model does not know everything before that date, and the label does not prove that every later fact is absent. Providers publish the date on the model card, and it marks where unaided recall becomes especially unsafe rather than a wall the model can see from inside.
Reach for it when deciding whether a question needs live retrieval. Anything that changes, such as prices, laws, officeholders, product versions, or today's incident status, should come from a current source even when its date appears older than the cutoff. Browsing and retrieved documents can supply newer facts in the request without retraining the base model.
Gotcha: asking the model whether it is current is not a freshness check. It may confidently invent a date or answer a recent question from an old pattern. Store the cutoff as model metadata, attach an as-of time to answers, and require citations for claims whose truth can change.
Ask AI for it
Add a freshness gate to this question-answering service. Store knowledgeCutoff in each model's config as an RFC 3339 date, classify every question for time-sensitive claims, and force retrieval whenever a claim may have changed or refers to an event after that date. Return answer, asOf, sources, and freshnessStatus under a JSON Schema; record each source URL and fetchedAt timestamp, reject citations fetched before the required asOf date, and answer 'Current information unavailable' when no qualifying source supports the claim.